Key Technology Of Real Time Person RE-Identification

The last several years has seen great development of computer vision and pattern recognition. As more and more new techniques has been applied to our common life, we have enjoyed more conveniences also with higher expectations. For example, with the spread of camera network in public occasions, public safety has called more and more attention nowadays. It becomes an increasingly urgent problem that we have to find a way to guarantee the public safety effectively by the videos made by the camera network. Person Re-identification is proposed to solve these problems. It aims to find the person from one or more cameras. We have proposed two re-identification methods, based on videos and image database respectively. My work includes:1. We have designed and implemented a reid system based on lots of videos. Our systems also follows the general framework of pattern recognition.(1) Unlike other researchers who will separate the object into numerous patches, we just divide the person into two parts with a ratio6:4, since our body complies with the law of the golden section;(2) we have designed a three-stage search strategy, at first, we filter the background by foreground extraction, and then we narrow the range further by dominant color matching, at last we will match the two zones by feature matching;(3) As to the scale problem, we introduced the priori information of the camera network, which will make our recognition more precisely;(4) We have tested our system in TRECVID dataset and our own videos, the recognition rate is about90%and we could process more than25frames per second, which satisfy our real-time usage;(5) We also have transmitted our system to the Hadoop platform which makes our system faster.2. We introduce the Re-rank to person re-identification field.(1) When we have got a simple rank of the datase by general methods eg. color histogram and LBP feature with Euclidean distance,we could use the information already known to rank again by RankSVM. This is somewhat like a mechanism of the search engine.(2) What we do will add more positive members into the query, which will improve the search result. The introduction of human-machine interactive information is also better able to meet the needs of intelligent processing and human-computer interaction in the future, which is the starting point of this thesis.